Svy: Subpop number observations dropping people

[Rachel Meadows]

Using NHANES and my subpopulation has 39,313 people but when I do a simple svy, subpop(if subpop==1): mean ridageyr the number of subpopulation observations is 37,425 instead of 39,313. I have no missing data on the survey parameter variables (MEC16YR, sdmvstra, sdmstra) or on ridageyr.

. svyset [pweight=MEC16YR], strata(sdmvstra) psu(sdmvpsu) vce(linearized) singleunit(centered)

pweight: MEC16YR
VCE: linearized
Single unit: centered
Strata 1: sdmvstra
SU 1: sdmvpsu
FPC 1: <zero>

. svy, subpop(if subpop==1): mean ridageyr
(running mean on estimation sample)

Survey: Mean estimation

Number of strata = 118 Number of obs = 82,091
Number of PSUs = 241 Population size = 1,415,698,832
Subpop. no. obs = 37,425
Subpop. size = 946,631,143.08
Design df = 123

--------------------------------------------------------------
| Linearized
| Mean Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
ridageyr | 47.36543 .2322556 46.90569 47.82516
--------------------------------------------------------------

Without weights or survey design variables, you can see there is no missing for age:

. mean ridageyr if subpop==1

Mean estimation Number of obs = 39,313

--------------------------------------------------------------
| Mean Std. Err. [95% Conf. Interval]
-------------+------------------------------------------------
ridageyr | 50.83229 .0932921 50.64944 51.01515
--------------------------------------------------------------

[Rachel Meadows]

Ok I figured it out with the help of STATA support.

My 1,888 missing people had '0' values for their MEC weight and that's why the svy, subpop code drops them out. People get a 0 value for the MEC exam weight if they didn’t go to the mobile examination at all. People who show up to the MEC exam, even if they don’t complete part of the exam, still get a nonzero weight value.

Since I will be doing multiple imputation on these missing values, I don't want these people to get dropped out. So I will use the interview weights, which have nonzero value weights for everyone and would then assume everyone showed up to the mobile exam which is what I’m artifically creating with the multiple imputation of those missing exam variables.